DeepHit
Deep Learning for Competing Risks · Also known as: Neural network competing risks, DL competing events
DeepHit is a deep neural network framework for survival analysis with competing risks. Introduced by Lee et al. in 2018, it extends DeepSurv to handle settings where multiple, mutually exclusive events can occur, such as disease-specific mortality versus death from other causes. DeepHit solves the challenge of personalized risk prediction when subjects can experience different types of terminal events, a common scenario in medical and reliability applications.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use DeepHit when you have multiple competing event types and want personalized risk estimates for each outcome. It is particularly valuable in clinical settings where patients might die from disease or from unrelated causes, in reliability engineering with multiple failure modes, or in customer analytics with multiple exit routes (churn reasons). The method assumes noninformative censoring and that the competing risks are defined and measurable. Prefer DeepHit over cause-specific hazard models when you expect complex interactions between covariates and event types.
Strengths & limitations
- Directly models cumulative incidence functions for each competing event, respecting probability constraints
- Captures complex covariate interactions and nonlinear effects for each event type simultaneously
- Produces calibrated, personalized risk estimates for all competing events from a single unified model
- Enables discovery of event-specific treatment effects without manual stratification
- Handles high-dimensional covariates and sparse event data more robustly than traditional approaches
- Requires sufficient event counts for each competing risk category to train reliably
- Network architecture and training objective add complexity compared to simpler competing risk methods
- Theoretical justification for the direct incidence modeling approach is less established than cause-specific models
- Computational cost is higher due to multi-task neural network training
- Interpretation of learned representations is challenging when multiple event types are modeled jointly
Frequently asked
How does DeepHit differ from modeling cause-specific hazards separately?
DeepHit directly models cumulative incidence functions using a unified neural network with multi-task learning. Traditional cause-specific approaches model each event's hazard independently, then mathematically combine them to get incidence. DeepHit's joint learning can capture event-specific covariate interactions better and more efficiently handles shared covariate effects.
What is the relationship between DeepHit and Fine-Gray regression?
Fine-Gray regression models subdistribution hazards using partial likelihood, assuming linearity. DeepHit replaces this linear assumption with a neural network, learning nonlinear relationships. Both target the same cumulative incidence function, but DeepHit offers greater flexibility.
How does DeepHit ensure that incidence probabilities sum to one?
The output layer uses a softmax-like constraint across event types to ensure all incidence curves are mutually exclusive and exhaustive. The training objective penalizes violations of this constraint, encouraging the model to respect the competing risks structure.
Can I use DeepHit if one event type is very rare?
Rare event types make training difficult. To handle imbalanced events, use weighted loss functions that upweight rare events during training, or stratified sampling to ensure each batch contains examples of all event types. Class balancing is critical for reliable predictions.
How should I choose the number of output heads for different event types?
Include one output head per distinct competing event type in your application. However, group rare event types together only if you do not need separate risk estimates. Test different groupings using cross-validation to maximize predictive accuracy.
What software implements DeepHit?
DeepHit implementations are available in pycox (Python), as well as custom TensorFlow and PyTorch code. Some packages built on lifelines also provide competing risks functionality, though specialized packages offer more sophisticated implementations.
Sources
- Lee, C., Zame, W., Yoon, J., & van der Schaar, M. (2018). DeepHit: A deep learning approach for dynamic survival analysis with competing risks. AAAI Conference on Artificial Intelligence, 32(1), 2314–2321. link ↗
- Fine, J. P., & Gray, R. J. (1999). A proportional hazards model for the subdistribution of a competing risk. Journal of the American Statistical Association, 94(446), 496–509. DOI: 10.1080/01621459.1999.10474144 ↗
- Katzman, J. L., et al. (2018). DeepSurv: Personalized treatment recommender system using a Cox proportional hazards deep neural network. Journal of Machine Learning Research, 40, 40–51. DOI: 10.1186/s12874-018-0482-1 ↗
How to cite this page
ScholarGate. (2026, June 3). Deep Learning for Competing Risks. ScholarGate. https://scholargate.app/en/survival/deephit
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
Compare side by side →